DualR

A clinical risk assessment web app that transforms medication history into phenotypic disease risk estimates using pre-computed drug associations from large language models and XGBoost inference.

Phenotypes: Type 2 Diabetes (T2D), Hypertension (HTN), Alcohol Use Disorder (AUD)

Su Lab · Biomedical Informatics, Biostatistics & Health Data Science · Indiana University School of Medicine

---

Repo Layout

DualR_app/
├── .github/workflows/
│   ├── ci.yml                 Runs on push/PR — builds frontend + backend
│   └── deploy-stage.yml       Manual deploy to stage EC2 (workflow_dispatch)
├── frontend/
│   ├── src/app.jsx            React SPA
│   ├── public/dr.png          Logo
│   ├── Dockerfile
│   ├── nginx.conf
│   ├── index.html
│   ├── main.jsx
│   ├── package.json
│   └── vite.config.js
├── backend/
│   ├── app/main.py            FastAPI prediction endpoint
│   ├── models/                XGBoost + parquet files (not in git)
│   ├── Dockerfile
│   └── requirements.txt
├── deploy/
│   ├── docker-compose.stage.yml
│   └── env.stage.example
├── docs/
│   ├── DEPLOYMENT.md
│   └── ARCHITECTURE.md
└── .gitignore

Run Locally

# Backend
cd backend
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
MODEL_DIR=models uvicorn app.main:app --reload --port 8000

# Frontend (separate terminal)
cd frontend
npm install
npm run dev        # proxies /api/* to localhost:8000

The frontend calls POST /api/predict on the backend. The backend loads XGBoost models and drug probability tables from backend/models/ at startup. Place the AoU model artifacts there before running.

Stage Deployment

Stage runs on a single EC2 instance (18.117.115.31, us-east-2). Deployment is manual:

1. Push code to main
2. Go to Actions → deploy-stage → Run workflow

The workflow runs docker compose ... up -d --build on the self-hosted runner on the EC2 instance.

See docs/DEPLOYMENT.md for setup details and docs/ARCHITECTURE.md for the system overview.

Secrets

• Never commit deploy/env.stage (gitignored)
• Model files go in backend/models/ on the EC2 instance — not in git
• No secrets belong in frontend code or VITE_* variables

Out of Scope (for now)

• Multi-environment deployment
• CDN, load balancer, auto-scaling
• LLM-backed scoring (vLLM endpoint is wired in the backend but not required for stage)
• Automated deploys on push (all deploys are manual for now)

License

Apache 2.0